FINDLOC: (was Implement simplification of minloc and maxloc)

FINDLOC: (was Implement simplification of minloc and maxloc)

[Darrell Reich via fortran]

PR54613 suggests MAXLOC could be adapted to implement FINDLOC by replacing max with any value. I took a look before it was simplified and the generated C code was over my head. Meanwhile, I am attempting to implement FINDLOC in Fortran.

Does this work on MAXLOC inspire you to be able to implement F08 FINDLOC? The spec is at link below for reference.

https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008 <https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008>
http://www.j3-fortran.org/doc/year/10/10-007.pdf <http://www.j3-fortran.org/doc/year/10/10-007.pdf>

Re: FINDLOC: (was Implement simplification of minloc and maxloc)

[Steve Kargl]

On Sat, Jan 06, 2018 at 11:28:46AM -0800, Darrell Reich via fortran wrote:
>
> PR54613 suggests MAXLOC could be adapted to implement FINDLOC
> by replacing max with any value. I took a look before it was
> simplified and the generated C code was over my head.  Meanwhile,
> I am attempting to implement FINDLOC in Fortran.
> 
> Does this work on MAXLOC inspire you to be able to implementr
> F08 FINDLOC?
> 

Hi Darrell,

Are you asking if someone else is working on findloc,
or are you asking for help in adding findloc to gfortran?

-- 
Steve

Re: FINDLOC: (was Implement simplification of minloc and maxloc)

[arrl via fortran]

On 1/6/2018 2:28 PM, Darrell Reich via fortran wrote:
> PR54613 suggests MAXLOC could be adapted to implement FINDLOC by replacing max with any value. I took a look before it was simplified and the generated C code was over my head. Meanwhile, I am attempting to implement FINDLOC in Fortran.
> 
> Does this work on MAXLOC inspire you to be able to implement F08 FINDLOC? The spec is at link below for reference.
> 
> https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008 <https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008>
> http://www.j3-fortran.org/doc/year/10/10-007.pdf <http://www.j3-fortran.org/doc/year/10/10-007.pdf>
> 
Do you mean Fortran such as
	i = findlocl(a(:) = val, .true., .false.)
....
    contains
     integer function findlocl(a,val,back)result(i)
     logical a(:),val,back
     integer i
     if(back)then
     do i=size(a),1,-1
       if(a(i).eqv.val)exit
       enddo
     else
     do i=1,size(a)
       if(a(i).eqv.val)exit
       enddo
       endif
       if(i>size(a))i=0
       return
     end function findloc
?
It's faster than ifort findloc.  The apparent advantage of a true
findloc implementation would be ability to terminate all loops early, if
that can be done without losing simd vector performance. In gfortran,
where little of this is vectorizable, this takes several times as long
as a plain search loop, due in part to the effort expended on creating a
new implicit array.
-- 
Tim Prince

Re: FINDLOC: (was Implement simplification of minloc and maxloc)

[Thomas Koenig]

Am 07.01.2018 um 13:27 schrieb arrl via fortran:
> On 1/6/2018 2:28 PM, Darrell Reich via fortran wrote:
>> PR54613 suggests MAXLOC could be adapted to implement FINDLOC by replacing max with any value. I took a look before it was simplified and the generated C code was over my head. Meanwhile, I am attempting to implement FINDLOC in Fortran.
>>
>> Does this work on MAXLOC inspire you to be able to implement F08 FINDLOC? The spec is at link below for reference.
>>
>> https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008 <https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008>
>> http://www.j3-fortran.org/doc/year/10/10-007.pdf <http://www.j3-fortran.org/doc/year/10/10-007.pdf>
>>
> Do you mean Fortran such as
> 	i = findlocl(a(:) = val, .true., .false.)
> ....
>      contains
>       integer function findlocl(a,val,back)result(i)
>       logical a(:),val,back
>       integer i
>       if(back)then
>       do i=size(a),1,-1
>         if(a(i).eqv.val)exit
>         enddo


That's one thing that I have been wondering about. Does it actually
make sense to implement FINDLOC by going backwards throught memory?
This is so much slower than going forward... of course, it
would depend on where the item to be found actually is.

Re: FINDLOC: (was Implement simplification of minloc and maxloc)

[Darrell Reich via fortran]

I am watching the recent discussions of "simplification of MAXLOC" and adding the BACK parameter.
I found the similar FINDLOC PR54613 in the database.
If the proposed solution to copy MAXLOC and add VALUE instead of MAX is valid then what is the level of effort to implement FINDLOC based on the recent improvements to MAXLOC?
Is it buy one get one free or a whole other project to dig into?

Having successfully built gfortran 8 and looking over the C code that implements MAXLOC, I am not in any position to contribute more than a test case at this point but I’m learning.

We’re upgrading from Fortran 77 to Modern Fortran 2008 / 2015 as we read up on the spec for these new intrinsic functions. It has been a long time since I took compiler design on a PDP-11. gcc project is very impressive.

>Are you asking if someone else is working on findloc, or are you asking for help in adding findloc to gfortran?

> PR54613 suggests MAXLOC could be adapted to implement FINDLOC
> by replacing max with any value. I took a look before it was
> simplified and the generated C code was over my head.  Meanwhile,
> I am attempting to implement FINDLOC in Fortran.
> 
> Does this work on MAXLOC inspire you to be able to implementr
> F08 FINDLOC?
> 

Re: FINDLOC: (was Implement simplification of minloc and maxloc)

[n8tm via fortran]

The standard requires the back option, which would be needed to find the last match.  In my test case, performance is very poor compared with a simple do loop.  If you're talking about hardware capability for buffering reverse memory access, the last cpu I saw lacking it was Athlon.  It's true the number of hardware supported reverse streams is expected to be small.

Sent via the Samsung Galaxy Tab E, an AT&T 4G LTE tablet
-------- Original message --------From: Thomas Koenig <tkoenig@netcologne.de> Date: 1/7/18  08:15  (GMT-05:00) To: tprince@intelretiree.com, fortran@gcc.gnu.org Subject: Re: FINDLOC: (was Implement simplification of minloc and maxloc) 
Am 07.01.2018 um 13:27 schrieb arrl via fortran:
> On 1/6/2018 2:28 PM, Darrell Reich via fortran wrote:
>> PR54613 suggests MAXLOC could be adapted to implement FINDLOC by replacing max with any value. I took a look before it was simplified and the generated C code was over my head. Meanwhile, I am attempting to implement FINDLOC in Fortran.
>>
>> Does this work on MAXLOC inspire you to be able to implement F08 FINDLOC? The spec is at link below for reference.
>>
>> https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008 <https://gcc.gnu.org/wiki/GFortranStandards#Fortran_2008>
>> http://www.j3-fortran.org/doc/year/10/10-007.pdf <http://www.j3-fortran.org/doc/year/10/10-007.pdf>
>>
> Do you mean Fortran such as
> 	i = findlocl(a(:) = val, .true., .false.)
> ....
>      contains
>       integer function findlocl(a,val,back)result(i)
>       logical a(:),val,back
>       integer i
>       if(back)then
>       do i=size(a),1,-1
>         if(a(i).eqv.val)exit
>         enddo


That's one thing that I have been wondering about. Does it actually
make sense to implement FINDLOC by going backwards throught memory?
This is so much slower than going forward... of course, it
would depend on where the item to be found actually is.